The Route Ahead Is Not the Traffic Now

TL;DR for operators

A routing system can observe congestion accurately and still price a route badly. The reason is temporal: current speed is highly relevant to a road a vehicle will enter now, but it can be a poor estimate for a segment the vehicle will not reach for several minutes.

HLSR¹ addresses that mismatch without continually reconsidering every vehicle in the network. It selects vehicles plausibly affected by detected congestion, then changes how much it trusts live versus predicted traffic according to when each route segment is expected to be reached.

The strongest evidence comes from a matched-scope ablation at 8,000 vehicles. Recommended HLSR reports a mean travel time of 380.6 seconds, while the same vehicle-selection scope using live traffic alone reports 438.1 seconds. The 57.5-second gap isolates the contribution of hybrid live-forecast route costing more cleanly than comparisons against entirely different routing systems.

For navigation and fleet operators, the design implication is narrower than “forecast more traffic.” Prediction is most valuable when attached to the timing of a specific routing decision, while intervention remains bounded. The paper demonstrates that logic in one fixed-seed SUMO environment, not in a live city deployment.

Current traffic ages before the vehicle reaches the road

Dynamic routing faces a basic timing problem. A vehicle chooses a path now, but that path contains segments reached at different future times. Treating the current measured speed on every segment as equally informative assumes that traffic conditions remain relevant throughout the trip.

HLSR instead calculates a horizon from estimated time to segment entry. Its route-costing speed is

with the weight on live traffic decreasing as the horizon grows:

The recommended parameters start the live-data weight at 0.75 and reduce it by 0.12 per horizon. The important design choice is not the particular decay schedule. It is that information is matched to when the vehicle will consume it: observations dominate nearby segments, while forecasts gain influence farther along the route.

That distinction explains why the paper’s most informative experiment is not a forecaster benchmark. It is the HLSR-LIVE ablation, which leaves the selective vehicle scope unchanged and removes forecast information from route costing.

Holding intervention scope fixed changes the interpretation

At 8,000 vehicles, HLSR records 380.6 seconds mean travel time and 0.61 reroutes per vehicle. HLSR-LIVE records 438.1 seconds and 0.80 reroutes.

This ablation is designed to isolate a controller component. Because the eligible rerouting set is held to the same optimized selection mechanism, the result cannot be explained simply by HLSR reconsidering more of the network. Removing hybrid costing makes trips 57.5 seconds slower, roughly a 15% increase relative to HLSR’s mean.

The experiment also weakens another intuitive explanation: that more frequent rerouting necessarily produces better outcomes. HLSR-LIVE reroutes more often yet performs worse. Under this simulated traffic process, reacting repeatedly to current conditions does not compensate for using information whose relevance decays before downstream segments are reached.

This is the paper’s strongest operational evidence because it separates two decisions that are often bundled together: which vehicles should be reconsidered and what traffic information should price their alternatives.

Selective intervention does not require giving up foresight

HLSR detects congestion using both occupancy and normalized velocity, then limits reconsideration to vehicles related to the bottleneck. Its selection mechanism includes a calibrated upstream depth and an additional two-hop expansion for vehicles approaching the affected area.

The purpose is bounded intervention. Vehicles with no plausible near-term relationship to the congestion event are not automatically replanned.

The main baseline comparison shows why this dimension deserves separate treatment from forecasting. At 8,000 vehicles, network-wide live travel-time Dijkstra, labeled CAIE-TT, reports a mean travel time of 408.1 seconds. Selective HLSR reports 380.6 seconds.

At higher demand, HLSR also has the lowest reported mean travel duration among the tested methods:

These are comparative simulation results, not evidence that selective routing universally beats network-wide optimization. Their narrower contribution is to show that intervention scope and predictive foresight are separable design dimensions. A system can constrain who gets rerouted while still using richer information about what those vehicles are likely to encounter.

Forecast accuracy is not the only forecast objective

Forecasting still matters inside HLSR, but the paper does not treat generic edge-level prediction accuracy as sufficient.

The forecasting backbone can be fine-tuned with an additional origin-destination path-ranking objective:

The encoder remains frozen while the output head is adapted. The added loss encourages predicted speeds to preserve the ordering of candidate paths, which is closer to the downstream routing decision than minimizing edge-level regression error alone.

The forecaster ablation tests this narrower proposition. Within the same HLSR stack, the per-edge LSTM produces 388.2 seconds mean travel time, 7.6 seconds slower than recommended HLSR. Huber-only LSTAN_GERPE reaches 400.8 seconds, 20.2 seconds slower.

These gains matter, but they are smaller than the 57.5-second penalty from removing hybrid live-forecast costing entirely. The forecasting model is therefore a supporting component, not the sole explanation for the system result.

Other leave-one-component-out tests point in the same direction. Replacing multi-objective route scoring with time-only scoring adds 25.0 seconds; removing the velocity component from congestion detection adds 19.7 seconds; removing driver personalization adds 12.4 seconds; and reducing the upstream depth from nine to four adds 5.1 seconds.

HLSR’s performance is a systems result produced by several linked decisions.

What navigation and fleet systems can take from this

For a navigation platform, the paper suggests a concrete allocation of engineering effort. The affected user is a driver or fleet vehicle approaching a known bottleneck. The decision is whether to retain the current route or allocate an alternative. The relevant condition is that downstream traffic conditions may change before the vehicle reaches them.

Under those conditions, Cognaptus infers three useful design principles.

First, prediction should be consumed selectively. Forecasts do not need to replace live telemetry everywhere. They become more relevant as the time between route calculation and segment entry increases.

Second, rerouting scope can be an operational control. Bounding intervention can reduce route-update volume and driver disruption while avoiding unnecessary network-wide behavioral changes.

Third, forecast models should be evaluated against the decision they support. If the operational output is route ranking, path-order preservation can matter independently of aggregate edge-level prediction error.

The paper also uses a multi-objective route cost combining normalized travel time, path length, route similarity, and occupancy balance. For fleet or mobility operators, this is a reminder that the cheapest predicted route is not necessarily the only system objective when repeated diversions can create new congestion or excessive route churn.

The evidence stops at the simulated controller

The architectural logic is clearer than the deployment claim.

The principal evaluation uses SUMO 1.18 on an OpenStreetMap-derived model of Tainan’s West Central District, with 204 intersections, 561 road segments, 20 origin-destination pairs, and demand files containing 8,000, 16,000, or 20,000 vehicles. Main comparisons use SUMO seed 42.

That design supports controlled comparisons among routing configurations. It does not establish how HLSR would behave across different cities, sensor-error regimes, traffic-generation processes, driver responses, or communications constraints. Several baselines are partially reimplemented on the authors’ TraCI platform, and the driver-personalization component has limited external validation.

The appropriate business takeaway is consequently architectural: test whether routing decisions improve when information freshness is matched to expected segment-entry time while intervention remains bounded. The reported travel-time superiority should be treated as a simulation result requiring replication across seeds, maps, demand patterns, and eventually live operations.

Route intelligence is partly a timing problem

HLSR’s most consequential result is not that traffic forecasts can improve routing. It is that the value of current versus predicted traffic changes along the route itself.

In the paper’s controlled comparison, leaving the rerouted-vehicle scope unchanged while replacing horizon-aware hybrid costing with live-only information raises mean travel time from 380.6 to 438.1 seconds. Broader replanning is therefore not the only lever available to a routing operator.

For navigation and fleet systems, that reframes the engineering choice. The system must decide who needs intervention, when current telemetry stops being the best representation of future road conditions, and whether its forecasting objective preserves the route choices that matter downstream.

HLSR provides a coherent simulated answer to those questions. Deployment evidence is still needed to determine how well that answer survives outside the simulator.

Cognaptus: Automate the Present, Incubate the Future.